Allow me to paraphrase his ask, which wasn't poorly stated in my opinion. He's not seen an AI DC before, but has some fairly recent general purpose DC experience. What are the differences? How've thing's changed?
Some models often report their training hardware. It doesn't take 5min to figure out that you can't "catch up" with it.
If you have the resources to "catch up", you would probably not ask poorly stated questions on HN, unless you have ulterior motives (which might be harmless, or not).
It seems like a question designed to steer people who might be unaware of such limitations into believing they can somehow buy cheap hardware and make their own datacenter, which is a weird propositon. That's an absurd idea, and such things would only be valuable for cheap spammy purposes (I've been investigating those schemes for a while).
So, we have a few scenarios here:
- Beginner guy who's too lazy to research for himself.
- Scammer trying to dump off hardware that was used for scams into other people's hands (my favorite kind of investigative meal).
- Someone else that is doing the same kind of investigative work I'm doing.
My answer suits these three profiles. It would humble the beginner, scare the shit out of the scammer, and elicit some respect from the investigator.
I know this is hard to follow, but keep focus on why these kinds of profiles are important. If you can think of another one, please share it, and I would rethink my answer in favor of that.
Dude. I don't know how what I asked and what you interpreted got so far apart, and yet I'm compelled to respond.
I've been in IT for more than 30 years. I've worked hands-on hardware for the first 20 of them. There's a datacenter in my office (though it's nearly empty as it's all been migrated to the cloud.)
What does a modern AI hosting datacenter look like? What's the network look like from a compute module to the perimeter of the DC? Is the hardware that hosts AI compute a 6U server with 20 RTX 4090's?
When Elon gets in trouble with having too many turbines...are they generating power at all times because it makes some kind of thermodynamic sense to convert LNG to power, rather than pulling power from the grid?
I'm not talking about the software used in training, I'm asking about the bare metal.
alganet|7 months ago
If you have the resources to "catch up", you would probably not ask poorly stated questions on HN, unless you have ulterior motives (which might be harmless, or not).
It seems like a question designed to steer people who might be unaware of such limitations into believing they can somehow buy cheap hardware and make their own datacenter, which is a weird propositon. That's an absurd idea, and such things would only be valuable for cheap spammy purposes (I've been investigating those schemes for a while).
So, we have a few scenarios here:
- Beginner guy who's too lazy to research for himself.
- Scammer trying to dump off hardware that was used for scams into other people's hands (my favorite kind of investigative meal).
- Someone else that is doing the same kind of investigative work I'm doing.
My answer suits these three profiles. It would humble the beginner, scare the shit out of the scammer, and elicit some respect from the investigator.
I know this is hard to follow, but keep focus on why these kinds of profiles are important. If you can think of another one, please share it, and I would rethink my answer in favor of that.
Damogran6|7 months ago
I've been in IT for more than 30 years. I've worked hands-on hardware for the first 20 of them. There's a datacenter in my office (though it's nearly empty as it's all been migrated to the cloud.)
What does a modern AI hosting datacenter look like? What's the network look like from a compute module to the perimeter of the DC? Is the hardware that hosts AI compute a 6U server with 20 RTX 4090's?
When Elon gets in trouble with having too many turbines...are they generating power at all times because it makes some kind of thermodynamic sense to convert LNG to power, rather than pulling power from the grid?
I'm not talking about the software used in training, I'm asking about the bare metal.